On 11.06.2013 14:37, Jonathan J. Helmus wrote:
Julian,
Since I am the author of the current percentile PR (https://github.com/numpy/numpy/pull/2970), I'm willing to try reimplementing percentile with the new partition functionality. I don't expect to have time to do this until the Scipy conference in two week, so if anyone else wants to give the implementation a try please feel free. Julian will you be at Scipy this year if I have any questions?
I wont be at the Scipy in June, but I can be reached via email or IRC (jtaylor on freenode and oftc) in the evening (UTC). btw. you don't need the minimum "trick" currently used in the branch for even element medians any more, so don't copy it for percentile. partition is specialized for small kth elements to be almost as fast as minimum, sometimes even faster if iterative partitioning is used (e.g. np.partition(d, (mid, mid + 1)))